Gemma 4 31B Strains Apple Silicon

The memory figures differ by precision and runtime path: Ollama’s library entry reflects the packaged model, Google’s card lists BF16 and Q4_0 sizes, and the reported idle footprint likely includes buffers and cache on top of the load itself. The practical takeaway for Apple Silicon users is that unified memory can exhaust quickly, so smaller quantizations, shorter contexts, or smaller Gemma 4 variants are the safer choice.